ADAPTIVE ENERGY-EFFICIENT FEDERATED DEEP LEARNING SYSTEM FOR CONTINUOUS MULTI-SENSOR HEALTH MONITORING IN EDGE ENVIRONMENTS

Abstract

The extensive growth of Internet of Medical Things (IoMT) devices has also made it possible to monitor health continuously, though traditional models have been struggling to maintain privacy, keep responsiveness in realtime, and conserve energy. Current methods tend to be associated with the high communication latency, low scalability to diverse sensors, and poor energy utilization, which limits their scalability to large-scale applications. To overcome these drawbacks, this study suggests an adaptive energy efficient federated deep learning architecture to health surveillance with multiple sensors under edge conditions. The framework integrates spatio-temporal transformer-based local model training, graph attention fusion for multi-sensor feature aggregation, and energy-aware federated optimization, allowing edge nodes to learn without sharing raw data. It is implemented in Python with the help of TensorFlow and PyTorch, using the ML-Based Health Monitoring IoMT Dataset of Kaggle. Experimental analysis has shown that the proposed model has an accuracy of 97.3% which is an improvement of about 3-5% compared to the baseline models and a better F1-score (96.8%) and precision (97.1%) and recall (96.4%). Moreover, it takes latency of 115 ms, and 9.6 inferences per joule, which represents real-time capability and optimization of resource. These findings confirm that attentionbased federated learning is effective in edge-based health monitoring. The proposed system presents a scalable, privacy-sensitive, and energy-efficient implementation of real-time remote healthcare with considerable advances in prediction accuracy and operational efficiency, timely detection of anomalies and emergency warnings, and is therefore extremely applicable to the implementation of heterogeneous IoMT settings.